VRDL_HW1: Bird Image Classification

This is homework 1 in NYCU Selected Topics in Visual Recognition using Deep Learning.

Installation

Requirements:

• CUDA version 10.1
• pytorch = 1.5.1
• torchvision = 0.6.1
• ml-collection
• scipy
• tensorboard

Or build the environment via:

$ conda env create -f environment.yml

Prepare Dataset

Put the train image in the train folder: ./dataset/train Put the test image in the test folder: ./dataset/test The dataset folder should be like:

./dataset
  |---classes.txt
  |---train_set_labels.txt
  |---val_set_labels.txt
  |---training_labels.txt
  |---testing_img_order.txt
  |---train
        |---xxxx.png
        |---xxxx.png
              .
              .
              .
  |---test
        |---xxxx.png
        |---xxxx.png
              .
              .
              .
  

Training Code

1. I follow the related work FFVT to use the Vision Transformer as my backbone model. The official pre-trained (on ImageNet21K) weight can be downloaded in https://github.com/jeonsworld/ViT-pytorch In my experiments, I use the weight of ViT-B_16.npz trained on ImageNet21K.
2. Train the model on the given Bird Dataset. I run our experiments on 2x1080Ti with bacth size of 4 for each card.

  $ python -m torch.distributed.launch --nproc_per_node 2 train.py --name [your exp name] --output_dir [your output dir] --dataset CUB_HW --model_type ViT-B_16 --pretrained_dir ./ViT-B_16.npz --img_size 448 --resize_size 600 --train_batch_size 4 --eval_batch_size 4 --learning_rate 0.02 --num_steps 10000 --eval_every 200 --feature_fusion

Evaluation code

Evaluate the model on the test data.

  $ python inference.py --name Test --output_dir Submission_check --dataset CUB_HW --model_type ViT-B_16 --pretrained_dir [path to your pre-trained model] --img_size 448 --resize_size 600 --eval_batch_size 4 --feature_fusion

Download Pretrained Models

Here is the model weight of my final submission. Please download the weights and run the above evaluation code.

• Final_submission_weights

Reference

My howework references the codes in the following repos. Thanks for thier works and sharing.

• ViT-Pytorch
• FFVT